VGGIN-Net: Deep Transfer Network for Imbalanced Breast Cancer Dataset.

Area of Science:

• Medical Imaging
• Computer Science
• Artificial Intelligence

Background:

• Breast cancer detection relies heavily on accurate image analysis.
• Deep learning models offer potential for improved diagnostic accuracy.
• Transfer learning can address challenges with smaller, imbalanced medical datasets.

Purpose of the Study:

• To develop a novel deep neural network architecture for breast cancer image classification.
• To leverage transfer learning from a large object dataset (ImageNet) to a smaller breast cancer dataset (BreakHis).
• To enhance classification performance through fine-tuning and regularization techniques.

Main Methods:

• A hybrid deep neural network architecture (VGGIN-Net) was designed, combining VGG16 and Inception modules.
• Transfer learning was employed by freezing VGG16 layers and integrating them with a new Inception module.
• Batch normalization, dropout, data augmentation, and block-wise fine-tuning were utilized to optimize the model.

Main Results:

• The VGGIN-Net model demonstrated significant improvements in classification performance.
• Fine-tuning the deep transfer network led to enhanced accuracy across different magnification factors.
• The proposed architecture achieved higher accuracies compared to existing state-of-the-art methods on the BreakHis dataset.

Conclusions:

• The VGGIN-Net architecture, incorporating transfer learning and fine-tuning, is highly effective for breast cancer image classification.
• The model's design facilitates effective transfer learning to other breast cancer datasets.
• This approach offers a promising solution for improving automated breast cancer diagnosis.